On this page
TL;DR:
- ROI is net benefits divided by full costs, times 100. The hard part is getting honest numbers to put in that sum.
- Time your chosen tasks before training. Asking staff later how much time they saved is unreliable, as one well-known study shows.
- Count every cost, especially staff time in the training room. In our made-up example it costs more than the provider's fee.
- Reduce the headline savings four times: how much work is really done with AI, how much is due to the training, how much freed time gets used, and how long the gain lasts.
- In the example, more real use improves ROI almost twice as much as a 25% discount on the fee.
Who this guide is for
This guide is for HR managers, training leads, finance teams and managers in Kenya who have paid for AI training, or are about to, and must answer one question: was it worth it?
You need a spreadsheet and some patience. You do not need a statistics degree. If you are still working out what to spend, start with our guide on how to budget for AI training.
What is ROI on AI training, in plain terms?
ROI compares what the training gave you in money with what it cost you in total. The standard formula is:
ROI (%) = (Benefits - Costs) ÷ Costs × 100
An ROI of 0% means you got your money back and nothing more. An ROI of 50% means that for every KES 1 you spent, you got your KES 1 back plus another 50 cents. A negative ROI means you lost money on the numbers you could measure.
Finance teams often like a second figure too, the benefit-cost ratio: Benefits ÷ Costs. A ratio of 1.2 means KES 1.20 back for every KES 1 spent.
The ROI Institute sets out Jack Phillips's method. It says a proper calculation needs two things that are easy to skip. Costs must be "fully loaded", meaning all of them. And you must isolate the effect of the programme from everything else that changed. It says that leaving out this step produces incorrect results.
What should you measure at each stage?
Measure in a chain, from how people felt to what changed in the business. Donald Kirkpatrick set out four levels. Phillips added a fifth, ROI.
| Level | Question | Cheap way to measure it | When |
|---|---|---|---|
| 1. Reaction | Was it relevant and useful? | Five-question form | Last hour of training |
| 2. Learning | Can they do it? | Each person improves a real prompt and catches an error in an AI draft | During training |
| 3. Application | Are they using it at work? | Team check-in, short pulse survey | Days 30, 60, 90 |
| 4. Impact | Did the work improve? | Re-time the same tasks, check error rates | Day 90 |
| 5. ROI | Was it worth the money? | The sum in this guide | Day 90, then month 12 |
A high satisfaction score is the weakest evidence on this list. People can love a training day and change nothing.
You may not need all five. The ROI Institute suggests that only about 5 to 10% of programmes justify a full ROI study, because each level costs more to measure. It suggests evaluating application for 30 to 50%. A team of 15 on a one-day course can stop at level 3. Use the full calculation when the spend is large or leadership has asked for a number.
Why can't you just ask staff how much time they saved?
Because people are poor judges of their own speed with AI. The clearest evidence comes from software developers, and the lesson applies widely.
In a 2025 randomised trial by the research group METR, 16 experienced developers worked on 246 real tasks. Before starting, they expected AI to speed them up by 24%. Afterwards, they believed it had sped them up by 20%. The measured result was that they took 19% longer.
Be careful with this study. It was small, it covered one profession, and it used early-2025 tools. METR's update of February 2026 says developers are now probably faster with AI than they were in early 2025. It also calls its own newer data only very weak evidence of how much faster. What holds up is the gap between what people felt and what the clock showed.
So use a clock where you can. Use surveys for adoption and confidence, and be wary of any survey that asks for hours saved.
Can you borrow someone else's percentage? Please don't. A study of 5,179 customer support agents found 14% more issues resolved per hour with an AI assistant, and 34% for novice workers. A Harvard and Boston Consulting Group experiment with 758 consultants found gains on some tasks and a drop in correct answers on another. These are useful for setting expectations. They measured access to AI on particular tasks. Neither measured the effect of a training course in your organisation.
How do you set a baseline before training?
Pick two to four tasks, and time them for two weeks before training starts. Without this step, every later claim is a guess.
Choose tasks that happen often, take real time and have a clear end. Routine letters, monthly reports and meeting minutes work well. Avoid rare or highly variable work.
Then record these five things for each task:
- Volume: how many times a month the team does it.
- Time: minutes per task, from a sample of at least 20 instances where possible.
- Quality: a simple score, or how many drafts come back for correction.
- Rework: minutes spent fixing errors.
- Cash spent on it: overtime, temporary staff, or outside suppliers.
Tell staff why you are timing the work, and that you are measuring the task and not judging them. Report at team level. Kenya's Office of the Data Protection Commissioner has a draft AI guidance note that is reported to list employee productivity monitoring among uses needing an impact assessment. It is only a draft, but tracking individuals is the wrong tool for this job anyway.
The best trick costs nothing: stagger the groups. Train group A in March and group B six weeks later. During those six weeks, group B is your comparison. If A's task time falls by 30% and B's falls by 5%, then something other than training caused that 5%. This is the most practical way to isolate the effect in a working organisation.
If you cannot stagger, ask the team leaders to estimate what share of the improvement came from training, and to list what else changed. Write down that it is an estimate.
What costs should you count?
Count everything the organisation gave up. The provider's invoice is only one line, and often not even the biggest.
| Cost | How to calculate it | Often forgotten? |
|---|---|---|
| Provider's fee | From the invoice | No |
| Staff time in training | People × hours × loaded hourly cost | Yes |
| Follow-up sessions | People × hours × loaded hourly cost | Yes |
| Organiser's time | Hours × loaded hourly cost | Yes |
| Measurement time | Hours × loaded hourly cost | Yes |
| Venue, lunch, travel, data bundles | Receipts | Sometimes |
| New tool licences | Only those bought because of the training | Sometimes |
Loaded hourly cost means what an hour of that person costs you, including employer contributions and benefits. Ask finance for an average by grade. Keep individual salaries out of the sheet.
A worked example in KES (all inputs are made up)
Every number in this example is invented to show the arithmetic. The company does not exist. The fee is not a FuKazee price or any market rate. The results are not from a real client. Replace everything with your own figures.
The scene: a made-up distribution company trains 40 finance and admin staff. It runs two days of training (14 hours) and two follow-up sessions (3 hours in total). Staff use a tool the company already pays for, so there are no new licences. Loaded cost is KES 700 an hour.
Step 1: add up the full cost
| Cost | Working | KES |
|---|---|---|
| Provider's fee (invented) | From the invoice | 400,000 |
| Staff time in training | 40 × 14 × 700 | 392,000 |
| Staff time in follow-up | 40 × 3 × 700 | 84,000 |
| Organiser's time | 30 hours × 700 | 21,000 |
| Measurement time | 20 hours × 700 | 14,000 |
| Room, lunch and data | Receipts | 40,000 |
| Total | 951,000 |
Staff time (KES 476,000 with follow-up) costs more than the fee. It is also the easiest line to leave out.
Step 2: work out the theoretical saving
Time includes checking and correcting the AI's work.
| Task | Volume a month | Minutes before | Minutes after | Hours saved if all done with AI |
|---|---|---|---|---|
| A. Routine emails and letters | 3,000 | 12 | 8 | 200 |
| B. Branch and department reports | 80 | 180 | 120 | 80 |
| C. Meeting notes and action lists | 240 | 35 | 20 | 60 |
| D. Spreadsheet clean-up | 160 | 90 | 60 | 80 |
| Total | 420 |
It is tempting to stop here. If you did, you would claim 420 hours × KES 700 × 12 months = KES 3,528,000. That is a return of 271%, and it would be wrong.
Step 3: apply four honest reductions
Reduction 1: how much of the work is really done with AI? At the 90-day check, the team reports these shares.
| Task | Share done with AI | Hours |
|---|---|---|
| A | 60% | 120 |
| B | 50% | 40 |
| C | 70% | 42 |
| D | 40% | 32 |
| Total | 234 |
Reduction 2: how much is due to the training? The staggered group shows that some gain would have come anyway. We credit training with 70%. That gives 234 × 0.70 = 163.8 hours.
Reduction 3: how much of the freed time is used well? Saved time is worth money only if it goes to useful work or reduces cost. We assume 70%. That gives 163.8 × 0.70 = 114.66 hours a month.
Reduction 4: how long does it last? We count 10 months of benefit in the first year, leaving out the first two while habits form.
Step 4: put a value on it
| Benefit | Working | KES |
|---|---|---|
| Time value | 114.66 hours × 700 × 10 months | 802,620 |
| Cash saved on overtime and temporary staff | 30,000 a month × 70% × 10 months | 210,000 |
| Total benefits | 1,012,620 |
Step 5: calculate
Net benefit: 1,012,620 - 951,000 = KES 61,620
ROI: 61,620 ÷ 951,000 × 100 = about 6.5%
Benefit-cost ratio: 1.06
That is a modest result, and it is far more believable than 271%. It also counts only one year. If people keep their habits, most of the training cost does not repeat.
Step 6: show a range
Give leadership three cases. A single number suggests more certainty than you have.
| Case | Hours really done with AI | Credited to training | Time used well | Cash saving per month | Total benefits (KES) | ROI |
|---|---|---|---|---|---|---|
| Cautious | 140 | 60% | 50% | None | 294,000 | -69% |
| Middle | 234 | 70% | 70% | 30,000 | 1,012,620 | 6.5% |
| Strong | 300 | 80% | 80% | 30,000 | 1,584,000 | 67% |
What the example teaches
Look at what moves the result. Start from the middle case.
- Win a 25% discount on the fee (KES 100,000 less): ROI rises from 6.5% to 19%.
- Get real use up from 234 to 300 hours, with no discount: ROI rises to 30%.
In this example, follow-up and manager support are worth more than hard bargaining. The cautious case also shows the real risk. If few people keep using what they learned, the money is largely gone. That is why training that doesn't stick is the most expensive kind.
What about the benefits you can't put a number on?
Report them separately, and never add them to the ROI figure. Mixing them in weakens the numbers you can defend.
Staff confidence, better first drafts, faster onboarding, fewer data mistakes and less after-hours work can all be real. List them with evidence: a quote, a survey result, a count. Phillips's framework calls these intangible benefits, and it treats them as part of the story.
Also report quality honestly. Take 20 outputs from before training and 20 from after. Remove the dates and ask a manager to score them without knowing which is which. A faster memo that returns twice for correction has not saved time.
And count the bad news. If AI caused an error that reached a client, or someone pasted personal data into the wrong tool, record it and its cost.
A simple measurement calendar
| When | What to do | Who |
|---|---|---|
| 4 weeks before | Agree the tasks and what success means | Sponsor, HR |
| 2 weeks before | Time the baseline | Team leaders |
| Last hour of training | Reaction form and one skills check | Trainer |
| Day 30 | Pulse survey: who is using what, and what is blocking them | HR |
| Day 60 | Team conversation; share what is working | Managers |
| Day 90 | Re-time the same tasks; score quality; run the sum | HR, finance |
| Month 12 | Update ROI with actual months | Finance |
Some drop in use after training is normal. What matters is where it settles. If use falls almost to nothing by day 60, you have an early warning while there is still time to act.
Three prompts you can copy
These use the four Cs we teach: Clarity, Context, Constraints and Customisation. Use team totals only. Leave out staff names and individual salaries.
Prompt 1: build the spreadsheet
I need an Excel model to calculate the ROI of a staff training programme. Context: I will enter tasks, monthly volume, minutes before, minutes after, share of work done with AI, share credited to training, share of freed time used well, loaded hourly cost in KES and months counted. Constraints: give me the column layout and the exact formula for each calculated column, then a cost table and a final ROI and benefit-cost ratio. Explain each formula in one sentence. Do not invent any numbers; use blank input cells. Customisation: I am comfortable with basic Excel but not pivot tables or macros.
Check every formula against the worked example above before you trust it. If the sheet does not give 6.5% with those inputs, something is wrong.
Prompt 2: write the 30-day pulse survey
Write a pulse survey of no more than six questions to send to staff 30 days after an AI training programme. Context: I want to know who is using what they learned, on which tasks, and what is getting in their way. Constraints: no question asks how many hours they saved; include one open question on what is blocking them; keep answers anonymous; it must take under three minutes. Customisation: staff will answer on their phones, so use simple English and short answer options.
Prompt 3: draft the report to leadership
Draft a one-page report to our CEO on the results of our AI training. Context: here are our measured results [paste team-level figures]. Constraints: start with the result in one sentence; separate measured savings, estimated savings and non-financial benefits; show cautious, middle and strong cases; state every assumption; include what did not work; do not add any figure that I have not given you. Customisation: the CEO reads quickly and trusts modest, well-supported numbers.
What goes wrong, and how to avoid it
| Mistake | Why it misleads | What to do |
|---|---|---|
| No baseline | Every later number is a guess | Time the tasks first |
| Reporting satisfaction as success | Enjoyment says nothing about use | Report levels 3 and 4 |
| Asking staff to estimate hours saved | Feeling faster and being faster can differ | Re-time real tasks |
| Timing only the AI's output | Checking and fixing take time too | Time the whole task |
| Leaving out staff time | Makes the cost look half its true size | Use the full cost table |
| Counting all saved hours as cash | Salaries are still paid | Apply the "used well" reduction |
| Crediting training with everything | New systems or staff also change results | Stagger groups, or state your estimate |
| Using a global percentage | Different tools, tasks and people | Use your own measurements |
| Adding soft benefits to the total | Damages trust in the hard numbers | Report them on a separate line |
When ROI is the wrong question
Some training is worth doing even when the sum is unclear. Teaching staff what not to paste into a public tool is a form of risk control. The value is in the incident that never happens, which no spreadsheet can show.
And training will not fix every result. If people have no access to the tool, if managers discourage its use, or if the task is blocked by a slow approval chain, a course will not produce a return. Find that out through our AI readiness assessment before you spend. Whether to use in-house or external trainers also changes your cost line, so model both.
Questions people ask
What is a good ROI for AI training?
We have not found a reliable published benchmark for AI training in Kenya, so be careful with anyone quoting one. Compare your result with what your organisation expects from other investments, and with what the same money could have done elsewhere.
How soon can we measure ROI?
Check use at 30 and 60 days. Run a first ROI estimate at 90 days, when habits have had time to form or fade. Update it at month 12 with actual figures.
Should we include tool licences in the cost?
Include licences bought because of the training. Leave out tools you were already paying for, and say so in the report. If your programme relies on free versions, the licence cost may be zero.
Is saved time real money?
Only partly. Salaries stay the same, so saved time becomes money when it reduces overtime or outside spending, or goes to work that earns or protects income. That is why the example applies a reduction, and shows cash savings on their own line.
How many tasks should we track?
Two to four. More than that and people stop recording properly. Choose tasks with high volume and a clear start and finish.
What if the ROI comes out negative?
Report it. Then look at where the chain broke: was it low use, weak manager support, the wrong tasks, or a programme that did not fit? A negative first-year figure with a clear diagnosis is useful to leadership. It may also mean the training was never likely to pay back in money, and was justified on risk or skills instead.
Sources
All links last checked 9 November 2026.
- ROI Institute, Comparison of Kirkpatrick and Phillips Evaluation Frameworks
- Kirkpatrick Partners, The Kirkpatrick Model
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
- METR, update on the developer productivity experiment (24 February 2026)
- Brynjolfsson, Li and Raymond, Generative AI at Work (NBER Working Paper 31161)
- Dell'Acqua and others, Navigating the Jagged Technological Frontier (working paper, September 2023)
- PPC Land, report on the impact assessment annex in the ODPC draft Guidance Note on AI
- Digital Policy Alert, status of the ODPC draft Guidance Note on Artificial Intelligence
Where FuKazee fits in
FuKazee trains teams in Kenya to use AI in their daily work. Since 2022 we have trained more than 3,000 people. Participants practise on their own kind of work, using the free version of each feature wherever one exists, so that they can keep going afterwards. They leave with reusable prompt libraries and recordings, and we check in with you after the training.
We cannot promise you a particular return, and you should be wary of anyone who does. We can help you choose tasks worth measuring before the programme begins. See how our programmes work on our AI training in Kenya page.
To talk through what a measurable programme would look like for your team, tell us the tasks you most want to improve and how many people do them. Email lead@fukazee.com.
Want guided support instead?
Plan AI training around your team's real work. Answer a few quick questions about your sector, team size and goals, and we will recommend the right programme and format. Sessions are built on your workflows and include a usage policy.
Rodney has run FuKazee's training since 2022 and still teaches inside every programme line: corporate sessions, the postgraduate cohort and the KCSE examiner webinars. First Class Honours in Journalism and Mass Communication, University of Nairobi.
